A data-intelligence framework for construction cost estimation under market uncertainty: conditional value and performance saturation

Accurate cost estimation in residential construction remains challenging under market uncertainty, particularly when projects are treated as homogeneous and market information is incorporated without assessing its marginal value. This study proposes an explainable data-intelligence framework combining project segmentation, shapley additive explanations (SHAP)-based variable ranking, baseline ablation and nested cross-validation in artificial neural networks (ANNs), support vector machine (SVM) and random forest (RF) to evaluate the incremental predictive value of market information. The analysis uses 372 residential projects developed in Tehran, Iran, between 1993 and 2008, with an interquartile range (IQR)-filtered sample for sensitivity analysis. K-medoids identifies two reasonably reproducible project-scale segments. Preliminary financial variables dominate baseline predictive information; when retained, market variables provide modest gains, whereas removing these proxies substantially increases the contribution of the first-ranked market indicators. Performance saturation is conditional on project segment, model class, baseline specification and stability tolerance, and is not detected in some configurations. No algorithm dominates across predictive accuracy, error, saturation and proportional agreement. The framework supports more selective use of market information in early construction cost estimation, subject to local recalibration and external validation.